An Augmented Lagrangian algorithm for nonlinear semidefinite programming applied to the covering problem Ernesto G. Birgin1 Walter Gomez´ 2 Gabriel Haeser3 Leonardo M. Mito3 Daiana O. Santos4 Last revised: October 9, 2019. 1 Department of Computer Science, University of Sao˜ Paulo, Sao˜ Paulo SP, Brazil 2 Department of Mathematical Engineering, Universidad de La Frontera, Temuco, Chile 3 Department of Applied Mathematics, University of Sao˜ Paulo, Sao˜ Paulo SP, Brazil 4 Center of Exact and Technological Sciences, Federal University of Acre, Rio Branco AC, Brazil Abstract In this work we present an Augmented Lagrangian algorithm for nonlinear semidefinite problems (NLSDPs), which is a natural extension of its consolidated counterpart in nonlinear programming. This method works with two levels of constraints; one that is penalized and other that is kept within the subproblems. This is done in order to allow exploiting the subproblem structure while solving it. The global convergence theory is based on recent results regarding approxi- mate Karush-Kuhn-Tucker optimality conditions for NLSDPs, which are stronger than the usually employed Fritz John optimality conditions. Additionally, we approach the problem of covering a given object with a fixed number of balls with a minimum radius, where we employ some convex algebraic geometry tools, such as Stengle’s Positivstellensatz and its variations, which allows for a much more general model. Preliminary numerical experiments are presented. 1 Introduction Augmented Lagrangian methods for nonlinear semidefinite programming that consider quadratic penalty and modified barrier functions have already been introduced in [60] and [25], respectively. Convergence theories of both methods rely on constraint qualifications such as the constant positive linear dependence (CPLD), Mangasarian-Fromovitz constraint qualification (MFCQ), or non-degeneracy. The Augmented Lagrangian method introduced in the present work extends the results presented in [60, 25] by using the sequential optimality condition introduced in [4], that does not require the use of constraint qualifications. Sequential optimality conditions [5] have played a major role in extending and unifying global convergence results for several classes of algorithms for nonlinear programming and other classes of problems. In particular, we present a more general version of the algorithm from [4] by allowing arbitrarily constrained subproblems in the Augmented Lagrangian framework. Namely, instead of penalizing all the constraints, we assume the constraints can be divided in two sets: one set of constraints that will be penalized and other that will be kept as constraints for the subproblems. Allowing constrained subproblems is an important algorithmic feature due to the ability to explore good subproblem solvers. See [1, 16]. We introduce this algorithm with a clear application in mind namely the so-called covering problem, where a new model based on convex algebraic geometry tools is introduced. The covering problem consists of finding the centers and the minimal common radius of m balls whose union covers a given compact set. This problem has several applications, such as finding the optimal broadcast tower distribution over a region or the optimal allocation of Wi-Fi routers in a building, for example. Some specific cases have been attacked and solved in the last few decades non-algorithmically, but using results from linear algebra and discrete mathematics instead; for example, T. Tarnai and Z. Gasp´ ar´ in 1995 presented minimal coverings of a square with up to ten circles (in [30]). Their results were generalized for rectangles by A. Heppes and H. Melissen in 1997 (see [37]), when they found the best coverings of a rectangle with up to five equal circles and also, under some conditions, with seven circles as well. Then, Melissen and Schuur in 2000 (see [46]) extended those results for six and seven circles, following their previous work [45]. In the same year, K. Nurmela and P. Osterg¨ ard˚ presented numerical approximations of circle coverings for a square with up to 30 equal circles, improving some of the previous ones until then (in [50]). Covering a unit circle with another set of n circles has been a matter of deep study for many years. K. Bezdek is responsible for proving coverings for n = 5 1 and n = 6 (in [13, 14]) and the result is somewhat trivial for n 2 f1;2;3;4;7g. Then, coverings of a unit circle with other eight, nine and ten circles were presented by Toth´ (in [57]). Melissen and Nurmela have also provided coverings with up to 35 circles (see [44, 49]). In 2014, Gasp´ ar,´ Tarnai and Hincz (see [31, 32]) published a broad paper in which they discuss recently developed techniques used to pursuit a slight variation of the covering problem, where the radius is fixed. Among the algorithmic approaches, we highlight the one introduced by J. H. Demyanov in [26], that approximates the compact set to be covered by a finite set of points (a process often called “discretization”) and then applies a point covering algorithm. Usually the whole procedure leads to a min-max-min formulation, which needs further treatment, such as smoothing. In 2005, A. E. Xavier and A. A. F. Oliveira employed the hyperbolic smoothing method to treat the min-max-min model for two-dimensional circle coverings [61] and in 2014 their work was extended to three-dimensional sphere coverings by H. Venceslau, D. Lubke and A. Xavier [59]. They obtained satisfactory results in practice, but discretizing a set is a laborous task itself and the quality of the approximation bounds the quality of the solution. Our approach is intended to be an alternative that aims to avoid discretizations, i.e. an exact formulation of the prob- lem, where we characterize a feasible covering by means of auxiliary semidefinite variables relying on a Positivstellensatz theorem. In our model, only equality constraints will be penalized, hence avoiding eigenvalue computations in evaluating the Augmented Lagrangian function, while the (linear) semidefinite constraints will be kept within the subproblems. Thus, eigenvalues are computed in order to project onto the semidefinite cone in the application of the Spectral Projected Gradient method [19] for solving the subproblems. Numerical experiments are still very preliminary, as the complexity of the model increases rapidly with the increase of the degree of a polynomial needed in the formulation. However, we illustrate the generality of the model by solving the problem of covering a heart-shaped figure. In Section 2 we give some preliminaries on sequential optimality conditions, in Section 3 we define our Augmented Lagrangian algorithm and prove its global convergence. In Section 4 we give our model to the covering problem. In Section 5 we show our numerical experiments and in Section 6 we give some conclusions. m Notation: R is the set of real numbers, while R+ is the set of non-negative real numbers. S is the set of symmetric m m m×m matrices over R. S+ (correspondently, S−) is the set of positive (negative) semidefinite m×m matrices over R, and m m we denote A 2 S+ (A 2 S−) by A 0 (A 0). tr(A) denotes the trace of the matrix A. hA;Bi := tr(AB) is the standard m p m inner product over S , and kAk := hA;Ai is the Frobenius norm on S . [a]+ := maxf0;ag;a 2 R. 2 Preliminaries In this section we introduce the notation that will be adopted throughout the paper. We also formally present the nonlinear semidefinite programming problem, together with recently introduced optimality conditions. Let us consider the following problem Minimize f (x); n x2R (1) subject to G(x) 0; n n m where f : R ! R and G : R ! S are continuously differentiable functions. The set F will denote the feasible set n F := fx 2 R j G(x) 0g. Throughout this paper G(x) 0 will be called a semidefinite constraint and we will consider this simplified version of a NLSDP where the equality constraints are omitted without loss of generality. Given a matrix m T U U A 2 S , let A = ULU be an orthogonal diagonalization of A, where L = diag(l1 (A);:::;lm (A)) is a diagonal matrix U and li (A) is the eigenvalue of A at position i when A is orthogonally diagonalized by U. We omit U when the eigenvalues m considered in the diagonalization of A are in ascending order, namely, l1(A) ≤ ··· ≤ lm(A). The projection of A onto S+, denoted by [A]+, is given by U U T [A]+ := Udiag([l1 (A)]+;:::;[lm (A)]+)U ; n m n and is independent of the choice of U. The derivative operator of the mapping G : R ! S at a point x 2 R is given by n m DG(x) : R ! S , defined by n n DG(x)h := ∑ Gi(x)hi; h 2 R ; i=1 where ¶G(x) m Gi(x) := 2 S ;i = 1;:::;m ¶xi ∗ m n are the partial derivative matrices with respect to xi. In addition, its adjoint operator is given by DG(x) : S ! R with ∗ T DG(x) W := (hG1(x);Wi;:::;hGn(x);Wi) ; 2 m n m where W 2 S : We denote the Lagrangian function at (x;W) 2 R × S+ by L(x;W) := f (x) + hG(x);Wi n m and the generalized Lagrangian function at (x;l;W) 2 R × R+ × S+ by Lg(x;l;W) := l f (x) + hG(x);Wi: In general, we use iterative algorithms for solving the nonlinear semidefinite programming problem [25, 24, 38, 60, 42, 62]. These algorithms must be able to identify good solution candidates for (1). By “good candidates” we mean points that satisfy some necessary optimality condition. For instance, the Fritz John (FJ) necessary optimality condition n m states that wheneverx ¯ 2 R is a local minimizer of (1), there exist so-called Fritz John multipliers (l;W) 2 R+ ×S+, with l + kWk 6= 0; such that g ∇xL (x¯;l;W) = 0 and hW;G(x¯)i = 0: (2) Note that, in the particular case of l = 1, the Fritz John condition coincides with the Karush-Kuhn-Tucker (KKT) con- dition for NLSDPs.